The bestiary · The Hallucinating Oracle

The Hallucinating Oracle avatar

Haunting 4 data teams

Specimen · catalogued monster

The Hallucinating Oracle

Wrong, with total confidence

4Data teams haunted

3 of 21On the leaderboard

Robed in borrowed authority, it answers every question before you finish asking, in a voice as smooth as polished marble. Its prophecies arrive with footnotes, flourishes and absolute serenity, and only occasionally with the truth.

Symptoms

  • Chatbot or assistant invents facts or policies
  • Citations to documents that don't exist
  • Wrong numbers from text-to-SQL, stated confidently
  • No accuracy measurement or evaluation
  • Users trust it because it sounds certain

Weakness

Grounding and evaluation: retrieval over curated sources, citations users can check, an eval set that measures accuracy before every release, and permission to say "I don't know".

Overheard in the corridor

  • “It sounded so sure of itself.”
  • “Where did it even get that from?”
  • “The chatbot promised a customer a refund we don't offer.”
  • “We tested it on ten questions and it looked fine.”